A Lightweight Transfer Learning Model for Face Mask Identification and Masked Face Recognition with Cross-Validation

Masked face recognition is vital for public safety, yet challenging due to occluded facial features and variations in pose, lighting, and mask type. In this study, we have conducted two distinct experiments. The first experiment focuses on detecting the existence of a face mask on a subject, referred to as Face Mask Identification (FMI). The second experiment aimed to identify the person while wearing a mask, referred to as Masked Face Recognition (MFR). We propose a lightweight transfer learning approach using EfficientNetB0 on two public datasets MDMFR and FMD. For the first experiment, we have trained and tested three models: CNN, ResNet50, and EfficientNetB0. Among them, EfficientNetB0 achieved 100% accuracy on both datasets. To evaluate the model’s generalization ability across datasets, we have conducted cross-dataset validation by training and validating on one dataset (FMD or MDMFR) and testing on the other, then reversing the roles of the datasets to ensure a comprehensive evaluation of performance. The results showed that EfficientNetB0 yielded the highest average accuracy of 97.00%, outperforming CNN (85.00%) and ResNet50 (94.50%). For the second experiment, conducted on MDMFR dataset, EfficientNetB0 achieved superior accuracy of 96.43% which is benchmarked with other state of the art works as well. Overall, this study demonstrates that due to its low parameter count and reduced training time, EfficientNetB0 is well-suited for deployment in resource-constrained environments.